Consideration is given to the piecewise linear neural network (PLNN), a neural network form of a piecewise linear classifier. The weight vectors of the cells in the middle layer of the network represent prototype patterns of the various classes to be recognized. The PLNN classifies input patterns by the minimum Euclidean distance to the prototype cells. The network structure is defined during training. The training is controlled by parameters ε and δ. If there is no cell of the correct class within distance ε of a training pattern, a new prototype cell is created. And all prototype cells within distance δ of the training pattern have their weights adjusted, either towards or away from the training pattern, depending upon their class. Cells of the same class that are moved close together are merged. The dynamics of the PLNN training is shown by examples of the prototype cells adapting to statistically generated data. The results of training backpropagation networks of similar size are described. The PLNN is also trained for the recognition of handwritten numerals. The processing of the image data, training of the network, and recognition results for PLNNs of various sizes are described. The results of training backpropagation networks of comparable size on the same data are also shown. The PLNN performance is generally superior, with far less training time required

A technique for the evaluation of a general continuous function f(x) is presented, and the design of an interpolating memory as an implementation of the technique is described. The technique partitions the domain of f(x) into segments and defines an interpolating (or approximating) function for each. The implementation is a memory subsystem that holds the parameters of the approximating functions and yields an interpolated function value on each read reference. Polynomial interpolating functions are considered in particular. Hardware requirements (memory and computational logic) are analyzed in terms of the required precision. It is shown that as long as f(x) has d+1 derivatives, where d is the degree of the interpolating polynomial, d+1 additional bits of precision of the computed f(x) are obtained for each additional address bit used in the interpolating memory. This establishes a tradeoff between memory and computational logic, which can be exploited in the design of a unit for a specific function, for any precision requirement. Furthermore, a single unit can be designed for any class of functions that have the required derivatives. Two examples of implementations for particular functions are presented.>

The phenomenon of repetition in the workload of a computer system is investigated from the point of view of its implications for systems design. Measurements taken from a CDC ***6000 Series interactive systems are presented. The first results show the distribution of repetitions of command executions (or program runs) over the classes of programs within the system. Then, a resource demand sequence is defined as the detailed pattern of hardware resources required by a program in execution. A technique for measuring the repetition of resource demand sequences is demonstrated, and the resulting measurements of this form of repetition are presented. The results show that the amount of repetition, though considerable, is insufficient to support the development of predictive schedulers. But they also indicate that a new approach to compilation – the use of an incremental recompiler – may effectively increase the productivity of an interactive computer system.

Certain queuemg discaphnes, such as processor sharing, the preemptive last-come-first-served discipline, and the infinite server queue, are known to result in network eqmltbnum state probabihties that have a convenient product form A generalization of the above dlsciphnes is introduced The general class is presented in the form of a parametenzed disciphne, called the last-batch-processor-sharing (LBPS) discipline The eqmhbrlum state probabilities for disciplines of the LBPS class are shown, and, by use of the concept of local balance, at ~s shown that arbitrary networks of LBPS queues have product form eqmlibrium state probabilities It as also shown that within the class of symmetric dlsciphnes, the LBPS form is necessary ff the product form solution ~s to be obtained for general service time dtstrlbutmns A dasctphne ~s symmetric ff the processor assignments to the customers in the queue depend on total queue occupancy and queue poslUon (relatwe arrival time) only Generahzations of the LBPS rule beyond the symmetric dlsciphnes are discussed A multiple customer-class form of the LBPS disclphne is also demonstrated, and it is shown to have the local balance property